[Bugfix] Reshape the dimensions of the input image embeddings in Qwen2VL (#9071)

2024-10-04 22:34:58 +08:00 · 2024-10-04 22:34:58 +08:00 · 3d826d2c52
commit 3d826d2c52
parent 0e36fd4909
1 changed files with 3 additions and 0 deletions
--- a/vllm/model_executor/models/qwen2_vl.py
+++ b/vllm/model_executor/models/qwen2_vl.py
@ -967,6 +967,9 @@ class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal,
                                           image_grid_thw=image_grid_thw)

        if image_embeds is not None:
+            image_embeds = self._validate_and_reshape_mm_tensor(
+                image_embeds, "image embeds")
+
            if not isinstance(image_embeds, torch.Tensor):
                raise ValueError("Incorrect type of image embeddings. "
                                 f"Got type: {type(image_embeds)}")